Treadmill and daily life

4 Treadmill and daily life Fall-related gait characteristics on the treadmill and in daily life, SM Rispens, JH van Dieën, KS van Schooten, LE Cofre Lizama, A Daffertshofer, PJ Beek, M Pijnappels, Journal of Biomechanics, submitted A summary of this work was presented at Dynamic Walking 2014 in Zurich, Switzerland (Rispens et al., 2014c). 51

Chapter 4 Abstract Background. Body-worn sensors allow assessment of gait characteristics that are predictive of fall risk, both when measured during treadmill walking and in daily life. The present study aimed to assess differences and associations between fall-related gait characteristics measured on a treadmill and in daily life. Methods. Trunk accelerations of 18 older adults (72.3 ± 4.5 years) were recorded during walking on a treadmill in the laboratory (DynaPort Hybrid sensor) and during daily life (DynaPort MoveMonitor). A comprehensive set of 32 fall-risk-related gait characteristics were estimated and compared between both settings. Results. For 25 gait characteristics a systematic difference between treadmill and daily-life measurements was found. Gait was more variable, less symmetric, and less stable during daily life. Fifteen characteristics showed a significant correlation between treadmill and daily-life measurements, including stride time and regularity (0.48 < r < 0.73; p < 0.023). No correlation between treadmill and daily-life measurements was found for stride time variability, acceleration range and sample entropy in vertical and medial-lateral direction, gait symmetry in vertical direction, and stability estimated as the Lyapunov exponent by Rosenstein s method in medial-lateral direction (r < 0.16; p > 0.25). Conclusion. Gait characteristics revealed less stable, less symmetric, and more variable gait during daily life than on a treadmill, while about half of the characteristics were significantly correlated between conditions. These results indicate that daily-life gait analysis is sensitive to static personal factors (i.e., physical capacity) as well as dynamic situational factors (i.e., behavior and environment), which may all represent determinants of fall risk. 52

Treadmill and daily life Introduction Quality of gait, as characterized among others by the stability, symmetry and variability of its kinematics, has been found to be associated with fall risk (Marschollek et al., 2011; Rispens et al., 2015b; Riva et al., 2013; Toebes et al., 2012; van Schooten et al., 2015; Weiss et al., 2013). In this context, gait quality is often assessed in the laboratory, under standardized conditions so as to minimize any variance that does not reflect the individual s walking ability, with treadmill walking at a fixed gait speed as the most controlled form. At the other extreme, gait quality is assessed in daily life, where outcomes may be influenced by behavioral and environmental factors that vary dynamically from situation to situation, in addition to the more static personal characteristics. In other words, laboratory assessment may for example reflect how stable someone can walk at a given speed under near-optimal conditions, whereas daily-life assessment might reflect how fast someone chooses to walk under the experienced environmental conditions. Viewed in this light, it is surprising that assessments in the laboratory (Riva et al., 2013; Toebes et al., 2012; Weiss et al., 2013) as well as in daily life (Marschollek et al., 2011; Rispens et al., 2015b; van Schooten et al., 2015; Weiss et al., 2013) result in gait characteristics that are predictive of fall risk. However, not for all characteristics do the findings of their relation with fall risk agree between the laboratory and daily life. For example, stride regularity, acceleration root-mean-square (RMS) and gait symmetry, all in medial-lateral (ML) direction, were found to discriminate between fallers and non-fallers in the laboratory, but not in daily life (Howcroft et al., 2013; Rispens et al., 2015a). While some of the gait characteristics used in fall risk prediction were reported to be correlated between laboratory and daily-life settings (Weiss et al., 2013), to our knowledge no comprehensive comparison between the two has been made to date. Such a comparison could facilitate the interpretation of differences and agreement found between predictive values of gait characteristics obtained in different settings with respect to fall risk. We therefore compared a comprehensive set of fall-related gait characteristics determined during treadmill and daily-life gait. We selected characteristics that had previously been shown to be associated with fall risk. Since we expected individual walking ability to influence outcomes in both cases, we hypothesized 53

Chapter 4 measures obtained in the two settings to contain common information reflected by positive correlations between settings. Furthermore, situational effects were expected to influence gait characteristics in daily life, and we hypothesized this to cause systematic differences, with gait in daily life being less stable, less symmetric and more variable. Methods Participants This study was part of the FARAO study, a cohort study on the predictive value for fall risk of gait characteristics derived from trunk accelerations in daily life. A subgroup of 18 older adults (72.3 ± 4.5 years, 11 males and 7 females, 1.73 ± 0.09 m; mean ± standard deviation) were invited for additional measurements in the laboratory during treadmill walking. Participants had a mini mental state examination score of 24 or higher and reported no musculoskeletal or neurological condition or medication that could have affected balance. All participants gave informed consent for both studies. The medical ethical committee of the VU University Medical Center Amsterdam (#2010/290, dailylife study) and the ethical committee of the Faculty of Human Movement Sciences, VU University Amsterdam (#2011/48, laboratory study), approved the study protocol. Protocol Daily-life accelerations were collected with a portable tri-axial accelerometer (DynaPort MoveMonitor, McRoberts, The Hague, The Netherlands), which was attached over the lumbar spine with an elastic belt around the waist. The accelerometer s range was set from -6 g to 6 g while its sampling frequency was set to 100 Hz. Participants were invited to wear the accelerometer for one whole week at all times, except during showering and other aquatic activities so as to avoid damage to the instrument. Trunk accelerations during treadmill walking were collected with a tri-axial inertial sensor, measuring accelerations and angular velocity (DynaPort Hybrid, McRoberts, The Hague, The Netherlands), of which only the accelerations were 54

Treadmill and daily life used in this study. The sensor s range and sampling frequency were equal to those in the daily-life measurements. The inertial sensor was attached in the same position as during daily-life measurements, i.e., over the lumbar spine with an elastic belt around the waist. Participants walked on the treadmill for 5 minutes at 1.2 m/s. As a safety precaution, participants wore a harness around their upper trunk, attached to the ceiling with a cable that provided enough slack not to interfere with walking. Data analysis A comprehensive set of gait characteristics were estimated from the daily-life data and from the treadmill data: stride time, stride time variability, stride regularity, acceleration RMS, acceleration range, gait symmetry (harmonic ratio), local dynamic stability (Lyapunov exponent), low frequency percentage, gait smoothness (index of harmonicity), dominant frequency s amplitude, and sample entropy. See Table 4.1 for references with algorithmic detail, and our previous studies (Rispens et al., 2014b; Rispens et al., 2015b) for general descriptions. All selected gait characteristics had previously been shown to be associated with past or future fall incidence or with clinical measures of fall risk in previous studies (Hamacher et al., 2011; Howcroft et al., 2013; Rispens et al., 2015a; Rispens et al., 2015b; van Schooten et al., 2015). For the treadmill data, the first and last minute were discarded to avoid any transients, leaving three minutes of gait data for analysis. For the daily-life data, we first identified the locomotion episodes that lasted at least 10 seconds, using a validated algorithm (Dijkstra et al., 2010) developed by the manufacturer of the sensors (McRoberts, The Hague, The Netherlands). For all data, gait characteristics were estimated from multiple 10-second epochs over which the median was taken. For treadmill walking, 3 minutes resulted in 18 epochs of 10 seconds. From daily-life gait episodes, we selected as many non-overlapping epochs of 10 seconds as possible, with any unused time equally divided over the beginning and end of the episode. In addition to the characteristics listed above, we estimated gait speed (Zijlstra and Hof, 2003), both during treadmill walking and daily life. Since speed was fixed during treadmill walking and gait speed may affect gait characteristics, we 55

Chapter 4 added a comparison of treadmill characteristics with speed-matched epochs in daily life. Daily-life epochs were selected if their estimated speed was between 90% and 110% of the estimated treadmill speed. The range was scaled to the estimated treadmill speed rather than the actual fixed treadmill speed, since we expected that a large part of the speed-estimate errors would be determined by individual differences in gait patterns. Statistics First, normality of the gait-characteristic estimates was confirmed by Kolomogorov-Smirnov test. Subsequently, we tested for systematic differences between treadmill and daily-life estimates. To this end, we applied a Wilcoxon s signed rank test since equality of variance between settings was frequently violated. Pearson correlations were used to assess to what extent treadmill and daily-life measurements provided common information. Significance of Pearson correlations was tested for the hypothesis that they were greater than zero, rather than different from zero. For each type of comparison, the significance level of 0.05 was adjusted using the Hochberg-Benjamini correction for multiple comparisons (Benjamini and Hochberg, 1995). Results The daily-life recordings contained, on average, 1755 (range 683 3328) epochs of 10 seconds of locomotion per participant, which decreased to 330 (range 74 814) when selecting only speed-matched epochs. Details about the estimated gait speeds on the treadmill and in daily life are provided in Figure 4.1. Twenty-five of the 32 gait characteristics displayed a systematic difference between treadmill and daily-life estimates (Table 4.1). Gait in daily life was characterized by larger variability (stride time variability and low frequency percentage < 0.7 Hz), less regularity (stride regularity and the dominant frequency s amplitude), less symmetry (harmonic ratio) and less local dynamic stability (higher Lyapunov exponent). No significant systematic difference between settings was found for estimates of acceleration RMS in vertical (VT) and anterior-posterior (AP) direction, acceleration range in VT and ML 56

Treadmill and daily life direction, low frequency percentage below 10 Hz in ML direction, the index of harmonicity in AP direction, and sample entropy in VT direction. 3 2.5 gait speed (m/s) 2 1.5 1 0.5 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 participant Figure 4.1: Box plot of gait speed estimates for 10-second epochs in daily life. For all individual participants, the ranges from minimum to maximum (vertical lines), the 25 th to 75 th percentiles (black rectangles) and the median (black horizontal lines) are shown. The grey rectangles indicate 90% to 110% of the treadmill speed estimated by the estimator used for daily-life data, which was the range used for the speed-matched analysis. For reference, the treadmill speed setting (1.2 m/s) is plotted as a horizontal grey line. For 15 of the characteristics we found a significant positive correlation between treadmill and daily-life estimates (Table 4.1). However, we did not find a significant correlation between treadmill and daily-life estimates for stride time variability, acceleration RMS and range in VT and ML directions, gait symmetry in VT and AP directions, local dynamic stability in terms of the largest Lyapunov exponent as estimated by Rosenstein s method, low frequency percentage under 10 Hz in ML direction, gait smoothness in ML and AP direction, the dominant frequency s amplitude in ML and AP directions, and sample entropy in VT and ML directions. 57

Chapter 4 Table 4.1: Mean and standard deviations (SD) of estimates based on treadmill and overall daily-life gait, Wilcoxon s signed rank test p-values and Pearson s correlation coefficient r (p-value) between settings. Significant values (p<0.039 for systematic differences, and p<0.023 for Pearson s correlation) are printed in bold face. Treadmill mean (SD) Daily-life mean (SD) Systematic difference p-value Pearson s correlation r (p) Stride Time (s) a 1.09 (0.07) 1.15 (0.11) 0.009 0.58 (0.006) Stride Time Variability (s) a 0.02 (0.00) 0.07 (0.04) <0.001 0.09 (0.362) Stride Regularity VT b 0.85 (0.07) 0.55 (0.18) <0.001 0.62 (0.003) Stride Regularity ML b 0.69 (0.09) 0.40 (0.14) <0.001 0.56 (0.008) Stride Regularity AP b 0.78 (0.07) 0.48 (0.16) <0.001 0.48 (0.021) Acceleration RMS VT (ms -2 ) c 2.32 (0.41) 2.24 (0.82) 0.640 0.45 (0.032) Acceleration RMS ML (ms -2 ) c 1.49 (0.26) 1.31 (0.27) 0.027 0.26 (0.146) Acceleration RMS AP (ms -2 ) c 1.45 (0.28) 1.46 (0.36) 0.829 0.73 (<0.001) Acceleration Range VT (ms -2 ) d 11.75 (2.68) 13.92 (4.00) 0.054 0.16 (0.264) Acceleration Range ML (ms -2 ) d 10.25 (2.04) 9.72 (2.39) 0.498-0.06 (0.600) Acceleration Range AP (ms -2 ) d 8.56 (1.97) 10.38 (2.94) 0.004 0.61 (0.004) Gait symmetry (Harmonic Ratio) VT c 2.80 (0.56) 1.71 (0.39) <0.001 0.14 (0.283) Gait symmetry (Harmonic Ratio) ML c 2.01 (0.41) 1.46 (0.19) <0.001 0.50 (0.016) Gait symmetry (Harmonic Ratio) AP c 2.53 (0.48) 1.48 (0.32) <0.001 0.32 (0.100) Lyapunov Exponent Wolf VT (s -1 ) a,e 0.83 (0.25) 1.42 (0.32) <0.001 0.59 (0.005) Lyapunov Exponent Wolf ML (s -1 ) a,e 1.31 (0.31) 1.73 (0.23) <0.001 0.51 (0.015) Lyapunov Exponent Wolf AP (s -1 ) a,e 1.02 (0.23) 1.62 (0.26) <0.001 0.55 (0.009) Lyapunov Exponent Rosenstein VT (s -1 ) a,f 0.61 (0.10) 0.76 (0.09) <0.001 0.28 (0.132) Lyapunov Exponent Rosenstein ML (s -1 ) a,f 0.57 (0.08) 0.64 (0.07) 0.008 0.01 (0.490) Lyapunov Exponent Rosenstein AP (s -1 ) a,f 0.54 (0.09) 0.65 (0.06) <0.001 0.21 (0.201) Low frequency percentage VT < 0.7 Hz (%) a 0.08 (0.09) 0.22 (0.15) <0.001 0.56 (0.008) Low frequency percentage ML < 10 Hz (%) a 85.18 (5.96) 86.70 (5.11) 0.308 0.39 (0.052) Low frequency percentage AP < 0.7 Hz (%) a 0.82 (0.38) 4.20 (2.99) <0.001 0.51 (0.015) Gait Smoothness (Index of Harmonicity) VT g 0.67 (0.10) 0.59 (0.11) 0.007 0.53 (0.013) Gait Smoothness (Index of Harmonicity) ML g 0.10 (0.09) 0.26 (0.17) <0.001 0.37 (0.068) Gait Smoothness (Index of Harmonicity) AP g 0.57 (0.10) 0.55 (0.10) 0.486 0.39 (0.057) Dominant Frequency s Amplitude VT d 0.82 (0.14) 0.66 (0.14) <0.001 0.65 (0.002) Dominant Frequency s Amplitude ML d 0.45 (0.12) 0.38 (0.09) 0.036 0.19 (0.220) Dominant Frequency s Amplitude AP d 0.63 (0.13) 0.49 (0.10) 0.002 0.20 (0.210) Sample Entropy VT h 0.26 (0.05) 0.24 (0.03) 0.092-0.09 (0.631) Sample Entropy ML h 0.40 (0.07) 0.35 (0.04) 0.025-0.00 (0.505) Sample Entropy AP h 0.29 (0.08) 0.25 (0.05) 0.027 0.61 (0.003) a (Rispens et al., 2014b; Rispens et al., 2015b), b (Moe-Nilssen and Helbostad, 2004), c (Menz et al., 2003), d (Weiss et al., 2013), e (Wolf et al., 1985), f (Rosenstein et al., 1993), g (Lamoth et al., 2002), h (Richman and Moorman, 2000) 58

Treadmill and daily life Table 4.2: Mean and standard deviations (SD) of estimates based on treadmill and speed-matched daily-life gait, Wilcoxon s signed rank test p-values and Pearson s correlation coefficient r (p-value) between settings. In daily life, only episodes with an estimated speed between 90% and 110% of the estimated treadmill speed were included. Significant values (p<0.038 for systematic differences, and p<0.020 for Pearson s correlation) are printed in bold face. Treadmill mean (SD) Speed-matched daily-life mean (SD) Systematic difference p-value Pearson s correlation r (p) Stride Time (s) a 1.09 (0.07) 1.12 (0.09) 0.019 0.80 (<0.001) Stride Time Variability (s) a 0.02 (0.00) 0.03 (0.01) <0.001-0.08 (0.624) Stride Regularity VT b 0.85 (0.07) 0.68 (0.12) <0.001 0.26 (0.153) Stride Regularity ML b 0.69 (0.09) 0.42 (0.12) <0.001 0.44 (0.033) Stride Regularity AP b 0.78 (0.07) 0.55 (0.15) <0.001 0.27 (0.143) Acceleration RMS VT (ms -2 ) c 2.32 (0.41) 2.42 (0.43) 0.140 0.80 (<0.001) Acceleration RMS ML (ms -2 ) c 1.49 (0.26) 1.34 (0.21) 0.030 0.38 (0.062) Acceleration RMS AP (ms -2 ) c 1.45 (0.28) 1.57 (0.25) <0.001 0.88 (<0.001) Acceleration Range VT (ms -2 ) d 11.75 (2.68) 13.78 (2.34) 0.010 0.31 (0.107) Acceleration Range ML (ms -2 ) d 10.25 (2.04) 10.70 (2.18) 0.510 0.08 (0.381) Acceleration Range AP (ms -2 ) d 8.56 (1.97) 10.94 (1.98) <0.001 0.71 (<0.001) Gait symmetry (Harmonic Ratio) VT c 2.80 (0.56) 2.14 (0.35) <0.001 0.20 (0.212) Gait symmetry (Harmonic Ratio) ML c 2.01 (0.41) 1.57 (0.23) <0.001 0.50 (0.018) Gait symmetry (Harmonic Ratio) AP c 2.53 (0.48) 1.79 (0.38) <0.001 0.11 (0.335) Lyapunov Exponent Wolf VT (s -1 ) a,e 0.83 (0.25) 1.24 (0.27) <0.001 0.46 (0.026) Lyapunov Exponent Wolf ML (s -1 ) a,e 1.31 (0.31) 1.74 (0.19) <0.001 0.52 (0.013) Lyapunov Exponent Wolf AP (s -1 ) a,e 1.02 (0.23) 1.51 (0.29) <0.001 0.44 (0.035) Lyapunov Exponent Rosenstein VT (s -1 ) a,f 0.61 (0.10) 0.77 (0.10) <0.001 0.31 (0.108) Lyapunov Exponent Rosenstein ML (s -1 ) a,f 0.57 (0.08) 0.62 (0.07) 0.056-0.00 (0.504) Lyapunov Exponent Rosenstein AP (s -1 ) a,f 0.54 (0.09) 0.62 (0.06) <0.001 0.37 (0.068) Low frequency percentage VT < 0.7 Hz (%) a 0.08 (0.09) 0.16 (0.11) <0.001 0.79 (<0.001) Low frequency percentage ML < 10 Hz (%) a 85.18 (5.96) 83.67 (6.12) 0.292 0.52 (0.013) Low frequency percentage AP < 0.7 Hz (%) a 0.82 (0.38) 2.63 (1.56) <0.001 0.12 (0.313) Gait Smoothness (Index of Harmonicity) VT g 0.67 (0.10) 0.70 (0.09) 0.188 0.58 (0.005) Gait Smoothness (Index of Harmonicity) ML g 0.10 (0.09) 0.22 (0.13) <0.001 0.52 (0.014) Gait Smoothness (Index of Harmonicity) AP g 0.57 (0.10) 0.59 (0.07) 0.286 0.56 (0.008) Dominant Frequency s Amplitude VT d 0.82 (0.14) 0.78 (0.13) 0.208 0.49 (0.019) Dominant Frequency s Amplitude ML d 0.45 (0.12) 0.31 (0.07) <0.001 0.44 (0.034) Dominant Frequency s Amplitude AP d 0.63 (0.13) 0.53 (0.12) 0.015 0.24 (0.173) Sample Entropy VT (REF) 0.26 (0.05) 0.23 (0.04) 0.029 0.24 (0.172) Sample Entropy ML 0.40 (0.07) 0.37 (0.05) 0.108 0.07 (0.392) Sample Entropy AP 0.29 (0.08) 0.24 (0.05) 0.003 0.66 (0.001) a (Rispens et al., 2014b; Rispens et al., 2015b), b (Moe-Nilssen and Helbostad, 2004), c (Menz et al., 2003), d (Weiss et al., 2013), e (Wolf et al., 1985), f (Rosenstein et al., 1993), g (Lamoth et al., 2002), h (Richman and Moorman, 2000) 59

Chapter 4 The speed-matched comparison revealed similar patterns in systematic differences and correlations between treadmill and daily-life gait characteristics as the overall comparison (Table 4.2). However, systematic differences were typically smaller and correlations were typically weaker for the speed-matched comparison, with the exception of the acceleration RMS and acceleration range, which had larger systematic differences and stronger correlations, and stride time, low-frequency percentages in VT and ML directions, gait smoothness, and sample entropy, which had stronger correlations. Discussion We compared gait characteristics estimated from measurements in daily life to those obtained under standardized laboratory conditions, i.e., during treadmill walking at a fixed speed. We found systematic differences between the two settings for most of the gait characteristics and found significant correlations between settings for about half of the gait characteristics. The systematic differences between settings may reflect an interaction of individual and environmental factors on the selection of instantaneous gait patterns, in terms of for example adapting speed, varying heading, or stepping over obstacles, which were controlled on the treadmill but not in daily life. An important factor influencing these findings might be the difference in gait speed between treadmill and daily-life settings. On the treadmill, gait speed was fixed at 1.2 m/s, close to the average preferred speed of a large cohort studied by Studenski and co-workers (Studenski et al., 2011), whereas it was on average lower and varied widely within and between our subjects in the daily-life measurements (as can be seen in Figure 4.1). Gait speed is known to have substantial effects on many gait parameters studied here (e.g., Bruijn et al., 2009a; Dingwell and Marin, 2006; Moe-Nilssen and Helbostad, 2004). We investigated this effect by comparing treadmill gait and daily-life gait with matched estimated speed. The systematic difference of treadmill gait with speed-matched daily-life gait was typically smaller than with all daily-life gait, which is in line with the assumption that part of the systematic difference is caused by speed differences. However, acceleration RMS and acceleration range showed larger systematic differences in the speed-matched comparison, 60

Treadmill and daily life which is surprising because these measures are strongly dependent on gait speed (Menz et al., 2003). We incorporated gait characteristics that had previously been shown to be associated with fall risk. It is therefore reasonable to assume that these reflect physical and cognitive capacities of the individual to generate a stable gait pattern, overcome perturbations and prevent falls. Many of the gait characteristics did show a significant correlation between laboratory and dailylife settings. We propose that the common information between the two settings is determined by personal factors (i.e., the individual s physical and cognitive capacity). This suggests that the characteristics that did not correlate significantly between settings are more strongly influenced by situational factors. These parameters will be discussed below. Stride time variability was not significantly correlated between settings. Stride time variability was solely found to be associated with fall risk when estimated under controlled conditions (Hausdorff et al., 2001; Toebes et al., 2012), and not when estimated in daily life (Rispens et al., 2015b; van Schooten et al., 2015). The association found under controlled conditions indicates that this parameter reflects an important aspect of the individual s capacity. Apparently this information is obscured in daily-life measurements due to the large effects of situational factors as suggested previously (van Schooten et al., 2014). Parameters indicating gait intensity, such as the RMS and range of the accelerations showed inconsistent correlations. RMS and range of ML and VT accelerations were not significantly correlated between settings, while in AP acceleration they were. The ML direction may predominantly reflect turning (van Schooten et al., 2015), which may have had a major influence on the ML intensity parameters in daily life, whereas turns were absent on the treadmill. Amplitude of movement in the VT direction is highly related to gait speed (Zijlstra and Hof, 2003), which was equal for subjects on the treadmill but not in daily life and which could therefore explain the absence of a correlation. This explanation is supported by the speed-matched analysis, in which we found a stronger correlation between settings for the acceleration RMS in VT direction (Table 4.2). The same dependency on gait speed holds for acceleration in the 61

Chapter 4 AP direction, but differences in sensitivity to spatiotemporal parameters such as step time and step length variability (Moe-Nilssen et al., 2010) might cause the differences between VT and AP directions of correlations found between treadmill and daily life. Another parameter showing different results for different directions was gait symmetry estimated by the harmonic ratio; correlations were non-significant in VT and AP directions, while gait symmetry in the ML direction was significantly correlated between settings. The non-significant correlation for gait symmetry in VT direction is surprising because previous studies, both in daily life and in the laboratory, revealed its association with fall risk (Doi et al., 2013; Rispens et al., 2015a; van Schooten et al., 2015). These latter findings were obtained during over-ground instead of treadmill walking. Since others did not find an association between fall risk and gait symmetry as determined from treadmill walking (Riva et al., 2013), this may indicate that treadmill walking strongly affects gait symmetry in VT and AP directions, rendering it less representative for personal fall risk factors. Local dynamic stability parameters as estimated by Rosenstein s method did not significantly correlate between treadmill and daily life, while those estimated by Wolf s method did. Although both methods have been designed to estimate the same concept, they apparently yield different estimates. Possibly the two estimators were affected differentially by other aspects of the measured signals than local dynamic stability. Further research is needed to address this issue. The low-frequency percentage below 10 Hz in the ML direction has been suggested to reflect the ability to quickly respond to balance disturbances (Rispens et al., 2015b), which may be called upon more frequently during daily life than on a treadmill. The different requirements between walking on a treadmill and in daily life might thus explain the lack of correlation between settings. Gait smoothness in ML and AP directions was not significantly correlated between treadmill and daily-life settings. After selecting only the speedmatched epochs in daily life, these parameters did show a significant correlation, indicating their gait-speed dependency. 62

Treadmill and daily life The dominant frequency s amplitude in ML direction, a measure of gait consistency, was not significantly correlated between laboratory and daily-life settings. As we found a significant correlation for stride regularity, which is estimated from the signal s auto-correlation and is similar to consistency, this raises the question how these two characteristics are different. An important difference is that stride regularity is based on stride time, whereas the dominant frequency in ML direction is not always the stride frequency but can also be an odd higher harmonic thereof. A closer look at the dominant frequencies in ML direction showed that on the treadmill they were on average typically between 4.2 and 5.1 Hz (14 of 18 participants), while in daily life they were typically below 3 Hz (12 of 18 participants). These differences in dominant frequencies may have had a significant impact on its amplitude, and may explain why we found no significant correlation between settings for this parameter. The last characteristic that did not show a significant correlation between treadmill and daily-life walking was sample entropy in VT and ML directions. For these directions, the studies on fall risk do not disagree on predictive value of treadmill versus daily-life gait (Rispens et al., 2015a; Riva et al., 2013). However, sample entropy in AP direction was correlated between the two settings, while in this case treadmill-gait estimates do discriminate fallers and non-fallers (Riva et al., 2013), while daily-life gait estimates do not (Rispens et al., 2015a). Correlations between gait characteristics on the treadmill and in daily life were typically lower after selecting the epochs in daily life that matched the treadmill in estimated speed (Table 4.2). The reduction of the number of epochs, caused by the selection, may have reduced the accuracy of the daily-life estimates, which would explain the lower correlations. As an exception, the acceleration RMS and range revealed stronger correlations between settings after matching the estimated speeds. Apparently, the effect of matching the estimated speeds is larger than the effect of the reduced accuracy for these characteristics possibly because their estimators are, like the gait-speed estimator that we used, highly dependent on the signal s amplitude. 63

Chapter 4 Although we were able to systematically compare gait characteristics obtained in a controlled setting with those obtained in daily life, this study does have some limitations. First, the estimation of gait characteristics from daily-life measurements may have caused substantial noise in our estimates. Gait characteristics in daily life were more variable over time than gait characteristics in the laboratory. To compensate for this considerably more data were obtained in daily life and only the median estimate was retained. We therefore do not expect this to have had a major influence on our results, given the reliability of the daily-life estimates shown previously (Rispens et al., 2015b). Furthermore, we used treadmill walking as the most controlled commonly used setting for observing gait. One may expect that other conditions such as over-ground walking or treadmill walking at preferred walking speed, which are less strictly controlled, would show better agreement with daily-life gait at least for some of the characteristics. For example, Dingwell et al. (2001) showed that gait variability and stability differ between over-ground and treadmill walking. As we pointed out in the discussion of symmetry measures, some of the measures might have been significantly correlated if we had used over-ground walking as our controlled condition. The effects of a specific constraint such as walking speed and the differences between over-ground walking in the laboratory and in daily life might be further uncovered if future research would include a condition of controlled over-ground walking and or treadmill walking at preferred or multiple walking speeds. Although our results suggest that daily-life gait analysis provides information on walking ability as well as behavioral and environmental influences on gait, and although all of these are likely to have a bearing on fall risk, it is not necessarily superior over gait assessment in the laboratory in prediction of fall risk. The results do indicate that a direct comparison of the predictive ability of parameters obtained in the laboratory and in daily life is warranted. Conclusion Estimates of fall-related gait characteristics from trunk accelerations typically displayed a systematic difference between measurements on a treadmill and in daily life. Half of the characteristics analyzed showed significant but at best 64

Treadmill and daily life moderate correlations between settings, whereas others were not significantly correlated. Our results indicate that daily-life gait is sensitive to static personal factors (i.e., physical capacity) as well as dynamic situational factors (i.e., behavior and environment), which may all represent determinants of fall risk. 65